AI on Bitsy Services Wiki

Prompt Engineering

Mon, 01 Jan 0001 00:00:00 +0000

Prompt engineering is the craft of wording an instruction to a language model so it produces the desired output reliably. Where context engineering decides what the model sees, prompt engineering decides how that instruction is phrased — role and system prompts, few-shot examples, chain-of-thought elicitation, output-format constraints, and the failure modes that come from ambiguous or over-stuffed prompts.

For a concrete instance — how the rules and the page request are worded for an agent that then acts on them — see the section’s running example, Claude Code: writing a page for this wiki.

Prompt Caching & Cost

Mon, 01 Jan 0001 00:00:00 +0000

Prompt caching lets a model reuse the processed form of a stable context prefix instead of recomputing it on every turn, turning the context ordering decisions — stable content first, volatile content last — directly into lower latency and spend. This page will cover token economics for agentic workloads: cache hit rate, what invalidates a cache, request batching, and model selection and routing.

For a concrete instance — why keeping CLAUDE.md and tool schemas stable at the front of the window makes a session cheap — see the section’s running example, Claude Code: writing a page for this wiki.

Agentic Workflows

Mon, 01 Jan 0001 00:00:00 +0000

An agentic workflow is a language model running in a loop with tools — planning, acting, observing the result, and deciding the next step — rather than answering in a single call. The loop is what makes it powerful and what makes it expensive: every iteration re-reads the accumulated history, every tool result enters the window, and every wrong turn compounds into the next. Getting value from an agent is mostly the discipline of not reaching for one until the problem needs it, then constraining it tightly when it does.

Agentic Engineering

Mon, 01 Jan 0001 00:00:00 +0000

Agentic engineering is the engineering discipline around systems where a model takes action: building, evaluating, deploying, and operating them. The other pages in this section name what happens inside a single turn or a single loop — prompt engineering wording instructions, context engineering curating what the model sees, agentic workflows shaping the loop. Agentic engineering is the lifecycle around them: the evals you write before launch, the traces you read after, the guardrails that keep the loop from cashing real money against a bad reasoning chain, the cost work that turns a working prototype into something a team can run on a budget.

MCP

Mon, 01 Jan 0001 00:00:00 +0000

The Model Context Protocol is the open specification for how a language model’s harness talks to an external tool server — connecting an agentic workflow to filesystems, APIs, databases, and other tools without bespoke per-integration code. MCP standardizes the wire format (JSON-RPC over stdio or HTTP), the tool-discovery handshake, and the shape of tool calls and results, so the same server works across any compliant harness.

This page is a stub. The protocol shows up in this section as the substrate beneath AI plugins — a Claude Code plugin frequently bundles one or more MCP servers, and the context-cost characteristics of MCP tools are what make plugin selection consequential.